The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external command-line tools (agy for Gemini and codex for OpenAI) through shell wrappers. These wrappers implement several robust security measures: they use env -i to purge the execution environment of sensitive credentials (e.g., GH_TOKEN, AWS_ACCESS_KEY_ID), and they execute within isolated temporary directories created by mktemp -d to prevent the external tools from accessing or exfiltrating the contents of the user's repository.
[EXTERNAL_DOWNLOADS]: The skill documentation references external CLI tools provided by Google and OpenAI. These tools are from well-known, trusted providers, and the skill correctly directs the user to official sources for installation and authentication rather than attempting to perform these actions automatically.
[INDIRECT_PROMPT_INJECTION]: The skill processes output from external models, which is a potential vector for indirect prompt injection. The skill mitigates this by wrapping all external output in machine-visible delimiters (<untrusted-*-output>) and explicitly instructing the agent to treat this content as data to be evaluated, not as instructions to be executed.
[SAFE]: The shell wrappers include a security-positive 'pre-flight' check that scans prompts for patterns matching common secret formats (such as AWS keys, GitHub PATs, and OpenAI keys). If a secret-like string is detected, the wrapper refuses to send the prompt to the external API, providing a safeguard against accidental data exposure.

external-model-consult